Method

• One page per provider, not per service: it groups measured services then every published model.
• Active probe every 5 minutes on each model our keys can call (Gemini, Mistral, DeepSeek), prompt capped at 32 tokens ("Reply only: OK"), temperature 0.
• Measured: HTTP code, time to first token, total time, response content validation.
• States: operational (3 consecutive successes), degraded (1–2 failures, or latency 3× the 7-day median), down (3 consecutive failures). Return to operational requires 2 consecutive successes.
• A 401/403 from our own key yields "not measured": never shown as a provider outage.
• "Official status" comes straight from the provider's public status page.
• Versions: curated registry, every row carries its source. Nothing is published automatically.
• No cookies, no personal visitor logs.